Cryptocurrency-predicting RNN Model

In this project, we will predict to buy the cryptocurrency or not after 3 minutes by analysing past 60 min cryptocurrencies price and volume using Recurrent Neural Network (RNN).

Introduction

The Dataset of Bitcoin, Bitcoin Cash, Ethereum and Litecoin Cryptocurrencies is avalable in crypto_data folder which will be used for training the RNN Network and the networks highest accuracy checkpoints wil be stored in the model director.

Network Architecture

Layer (type)                 Output Shape              Param #
=================================================================
cu_dnnlstm (CuDNNLSTM)       (None, 60, 128)           70656
_________________________________________________________________
dropout (Dropout)            (None, 60, 128)           0
_________________________________________________________________
batch_normalization (BatchNo (None, 60, 128)           512
_________________________________________________________________
cu_dnnlstm_1 (CuDNNLSTM)     (None, 60, 128)           132096
_________________________________________________________________
dropout_1 (Dropout)          (None, 60, 128)           0
_________________________________________________________________
batch_normalization_1 (Batch (None, 60, 128)           512
_________________________________________________________________
cu_dnnlstm_2 (CuDNNLSTM)     (None, 128)               132096
_________________________________________________________________
dropout_2 (Dropout)          (None, 128)               0
_________________________________________________________________
batch_normalization_2 (Batch (None, 128)               512
_________________________________________________________________
dense (Dense)                (None, 32)                4128
_________________________________________________________________
dropout_3 (Dropout)          (None, 32)                0
_________________________________________________________________
dense_1 (Dense)              (None, 2)                 66
=================================================================
Total params: 340,578
Trainable params: 339,810
Non-trainable params: 768

CuDNNLSTM is LSTM layer with CUDA gpu support which makes it much more faster then normal LSTM layer

Result

Train Accurecy after 20 epoch: 74.45%

Test Accurecy after 20 epoch: 54.32%

Credits

Special thanks to Harrison Kinstey for pythonprogramming.net lectures.